The story started around ten years ago, when studies of silver sulphide systems at Japan’s National Institute of Materials Science (NIMS) revealed unexpected behaviour that mimicked the synapses linking neurons to relay messages around the nervous system. Under an applied electric potential, silver ions were reduced and formed a conducting silver filament that disintegrated when the potential was removed – an "atomic switch". Applying electric pulses in repeated patterns gave responses that depended on the system’s history – characteristic of the long- and short-term memory in synapses.

"A lot of people are now embedding these 'artificial synapses' into circuits," says Adam Stieg, at the University of California in LA (UCLA). "But in the brain there’s not just one synapse, there are billions." After a little brainstorming, he and James K Gimzewski, who is affiliated with NIMS as well as UCLA and the University of Bristol, UK, looked for ways of producing networks containing numbers of artificial synapses comparable to those in a brain.

For these quantities pure top-down precision engineering approaches become untenable. Alongside colleagues at UCLA and NIMS, Stieg and Gimzewski also tried creating self-assembled electrochemically grown silver nanowires through simple deposition processes with no applied field. They then functionalised the nanowires grown through this "electroless deposition" to form atomic switches at junctions between nanowires. But, while easier to fabricate, the random and freely grown nanowires lacked a reliable network density. The winning strategy combined both top-down and bottom-up procedures following a design approach the researchers describe as nanoarchitectonics. This design approach is at the heart of work at the NIMS International Center for Materials Nanoarchitectonics (MANA) led by Masakazu Aono, also a collaborator in this research.

The team grew atomic switch elements over a lithographically patterned multielectrode array, which provided a way to investigate the system at a macroscopic level with traditional electronic measuring techniques. "You don’t need intimate details at every level," Stieg tells "You can look for patterns of activity."

Adam Stieg gives a 5 minute talk on complexity in nanotechnology research at UCLA July 2013.

Making sense of complexity

The researchers found that where “switching” occurred at a junction - i.e. a filament formed - a potential drop across the filament affected the potential distribution across the network. The result is a complex time-varying network of interacting synaptic components that respond to stimuli, but not always in the same way. Because of the artificial synapses at the junctions of the nanowires, the networks have a memory component to their behaviour as well.

"The properties that you see emerging are ubiquitous in nature," says Stieg. "I’m constantly amazed that we have reasonable levels of success with these systems – that these systems make responses that make sense."

While cautious not to overstate the potential of this kind of system, Stieg points out that it could be interesting for the development of alternative computing architectures. "Traditional computing has limitations and while there is no one panacea, these atomic switch networks might be useful for alternative approaches like reservoir computing that are typically better at tasks like pattern recognition."

In reservoir computing, inputs feed into a dynamic system or "reservoir", which maps the input to a higher dimension (like the outputs in the nonlinear transformation figure above) for computations that more closely resemble neural processes. Artificial neural networks also offer more brain-like computation and progress has been made by developing new algorithms. Stieg and his team, on the other hand, are not using algorithms - the neural network computing characteristics are implicit in the hardware of their systems.

"We have a saying: if it does what you expect it to do it’s impressive but not interesting - algorithmic systems generally do what they’re designed to do," says Stieg. "There’s no software in your brain but you can learn and adapt readily - it makes it a bit messier but it gives the capacity to perform a range of tasks."

Details of the research are reported in Nanotechnology 26 204003, part of the focus collection on Creating Novel Materials on the Basis of Mathematics and Nanoarchitectonics.